Anthropic released Claude Fable 5.2 and the restricted Claude Mythos 5.2 on Monday, claiming a substantial lead over a model identified as GPT-6 Cyber in a cybersecurity benchmark. The company says the new models improve coding, reasoning and vision, while generating responses up to 30% faster than their predecessors.
The headline comparison requires an important qualification: Anthropic’s published table reports a 76.3% score for Mythos 5.2 on CyScenarioBench, against 61% for GPT-6 Cyber. Although the results appear on the Claude Fable 5.2 product page, the cybersecurity result is explicitly attributed to Mythos 5.2, not the generally available Fable configuration. Anthropic’s page does not identify GPT-6 Cyber beyond that label or establish whether it is an OpenAI product.
Anthropic also reports that Mythos 5.2 achieved 89.6% on its CVE Rediscovery evaluation, compared with 78.3% for Mythos 5.1. Those figures are vendor-published model evaluations, not evidence that the system has exploited vulnerabilities in live environments. The company provides no CVE identifier, exploit-in-the-wild report, public proof of concept or threat-actor attribution associated with the release.
Access controls reflect the same dual-use concern behind the benchmark. Fable 5.2 is available to customers through Anthropic’s services and cloud marketplaces at $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for version 5.1. Anthropic says its cybersecurity safeguards can block or limit advanced requests and route flagged queries to less capable models, while organizations conducting legitimate security research can apply for verification.
Mythos 5.2, which Anthropic describes as sharing Fable 5.2’s underlying capabilities, is reserved for organizations approved through programs such as the Cyber Verification Program. The platform documentation lists a one-million-token context window and a maximum output of 128,000 tokens for both models, capabilities aimed at long-running agentic work as well as security analysis.
Security teams should therefore treat the announcement as a model-capability claim rather than a new vulnerability disclosure. Anthropic says its benchmark scores use each model’s highest available effort and that CyScenarioBench results are based on unique tasks for each lab; independent reproduction, task-level results and fuller methodology will be needed to determine how meaningful the reported advantage is in operational penetration testing or defensive work.
